A Hybrid CNN for High-Accuracy Document Skew Detection and Correction

Skew detection and correction are critical pre-processing steps for robust Optical Character Recognition (OCR) and document analysis systems. Traditional methods, such as those based on the Projection Profile or Hough Transform, often fail on documents with complex layouts, non-textual content, or low-resolution scans. In this paper, we propose SkewNet, a novel Convolutional Neural Network (CNN) architecture designed for fast and highly accurate skew angle estimation. Skew Net employs a hybrid prediction head that combines the strengths of both classification and regression. A classification head predicts a coarse angle from a set of discrete bins, while a regression head predicts a fine residual offset, allowing for sub-degree precision. We train our model on a large-scale synthetically generated dataset and evaluate it on a diverse public benchmark. Experimental results show that SkewNet achieves a Mean Absolute Error (MAE) of just 0.08 degrees and an accuracy of 99.8% within a 0.2- degree tolerance, outperforming both traditional methods and recent deep learning-based approaches. Furthermore, our lightweight model design achieves an inference speed of 5ms per image on a GPU, making it suitable for high-throughput applications.